An electric vehicle operation safety bench detection evaluation method and system

Abstract

The application discloses an electric vehicle operation safety bench detection evaluation method and system, and belongs to the technical field of automobile safety detection. The evaluation method comprises the following steps: performing feature extraction on the aligned original visual data and original bench data, performing safety evaluation processing on the fused features, obtaining a comprehensive safety index, performing safety judgment processing according to the comprehensive safety index and a preset global threshold, and obtaining a judgment result. The present application fuses visual perception technology and traditional bench data, improves the precision of electric vehicle operation safety detection, can more comprehensively and accurately evaluate the safety state of the vehicle under different working conditions, enhances the stability and adaptability of detection through a space-time graph convolution network, can cope with complex and changeable test environments and vehicle operation states, and provides an efficient and reliable solution for the safety detection of electric vehicles.

Description

Technical Field

[0001] This application belongs to the field of automotive safety testing technology, and in particular relates to a bench test and evaluation method and system for the operational safety of electric vehicles. Background Technology

[0002] With the rapid growth in the number of electric vehicles, their operational safety issues are becoming increasingly prominent. Among these, malfunctions such as abnormal drive systems, brake system failures, and battery thermal runaway can lead to serious traffic accidents. Current electric vehicle operational safety testing primarily relies on traditional bench testing, collecting data such as vehicle speed, acceleration, and pedal opening via the CAN bus, and combining this data with preset operating conditions to evaluate drive-braking performance. For example, the bench testing method proposed in patent CN 117268793 A can extract motion features through principal component analysis, but it relies solely on traditional sensor data and lacks direct observation of the vehicle component status. On the other hand, while patent CN 119135725 A achieves real-time acquisition and analysis of bench data, it does not involve visual perception technology. For instance, in the event of a brake system failure, bench data may only reflect abnormal deceleration, while visual technology can directly observe brake disc wear or localized overheating. However, existing solutions lack a joint analysis mechanism for visual and bench data, leading to delays in hazard identification.

[0003] Therefore, it is necessary to provide a new bench test and evaluation method and system for electric vehicle operation safety to solve the above-mentioned technical problems. Summary of the Invention

[0004] The purpose of this disclosure is to provide a bench test and evaluation method and system for electric vehicle operation safety in order to solve the above-mentioned problems.

[0005] This disclosure achieves the above objectives through the following technical solutions: A bench test and evaluation method for the operational safety of electric vehicles includes the following steps: The acquired raw visual data and raw bench data are aligned to obtain synchronized visual data and bench data; the visual data includes wheel rotation trajectory, vehicle pitch angle and brake disc temperature distribution, and the bench data includes vehicle speed, acceleration and brake pedal opening. Feature extraction is performed on the visual data to obtain motion state features and component state features. Feature extraction is also performed on the bench data according to a preset autoencoder network to obtain drive segment feature parameters and braking segment feature parameters. The motion state features include wheel speed change rate and vehicle pitch angle change parameters. The component state features include the proportion of high-temperature pixels. The motion state features, component state features, driving segment feature parameters, and braking segment feature parameters are fused to obtain a spatiotemporal feature fusion matrix. The spatiotemporal feature fusion matrix is ​​then input into a preset spatiotemporal graph convolutional network for safety assessment and outputs a comprehensive safety index. Based on the comprehensive safety index and the preset global threshold, a safety assessment is performed to obtain the assessment result, which is used to adjust the bench test parameters.

[0006] As a further optimization of this disclosure, the acquired raw visual data and raw test bench data are aligned to obtain synchronized visual data and test bench data, including: A synchronous trigger pulse signal is generated through the FPGA hardware triggering mechanism. The synchronous trigger pulse signal is used to control the vision acquisition system and the CAN bus acquisition device to start synchronously and acquire raw vision data and raw bench data respectively. Based on the internal clock of the vision acquisition system and the clock of the CAN bus acquisition device, the acquired raw vision data and raw bench data are timestamped to obtain an aligned timestamp sequence. Based on the aligned timestamp sequence, the original visual data and original test bench data are resampled using cubic spline interpolation to obtain synchronized visual data and test bench data.

[0007] As a further optimization of this disclosure, feature extraction is performed on the visual data to obtain motion state features and component state features, including: Based on the image data corresponding to the wheel rotation trajectory, the pixel displacement difference between adjacent frames is calculated using the optical flow method, and the optical flow equation is solved using the Lucas-Kanade algorithm to obtain the wheel rotation speed change rate. Based on the binocular vision data corresponding to the vehicle pitch angle, the coordinates of the preset marker points on the vehicle are matched by a three-dimensional reconstruction algorithm, and the vehicle pitch angle change parameters are output. Based on the brake disc temperature distribution, the U-Net segmentation network is used to segment the infrared image corresponding to the brake disc temperature distribution, calculate the pixel ratio of abnormal high temperature areas, and obtain the high temperature pixel ratio.

[0008] As a further optimization of this disclosure, feature extraction is performed on the test bench data based on a preset autoencoder network to obtain drive segment feature parameters and braking segment feature parameters, including: An autoencoder network based on gated recurrent units and temporal attention mechanism is constructed. The encoder of the autoencoder network includes an LSTM unit, a GRU gate module, and an attention weight calculation module. Based on vehicle speed and acceleration, the GRU module and the LSTM unit are sequentially used to perform temporal correlation filtering and time series encoding to obtain the LSTM hidden state sequence. The attention weight calculation module is then used to weight the LSTM hidden state sequence to obtain the weighted driving temporal features. Time series analysis is performed on the weighted driving time series features to calculate the acceleration change rate and speed fluctuation index, and driving segment feature parameters are obtained. The opening gradient sequence is calculated based on the brake pedal opening, and the time series correlation filtering and gradient encoding are performed sequentially through the GRU module and the LSTM unit to obtain the LSTM gradient hidden state sequence. The attention weight calculation module is used to weight the LSTM gradient hidden state sequence to obtain the weighted braking gradient features. Gradient analysis is performed on the weighted braking gradient characteristics to calculate the braking response delay and opening change gradient, thereby obtaining the braking segment characteristic parameters.

[0009] As a further optimization of this disclosure, the motion state features, component state features, driving segment feature parameters, and braking segment feature parameters are subjected to feature fusion processing to obtain a spatiotemporal feature fusion matrix. This spatiotemporal feature fusion matrix is ​​then input into a preset spatiotemporal graph convolutional network for security assessment processing, outputting a comprehensive security index, including: Based on the motion state features, component state features, driving segment feature parameters, and braking segment feature parameters, a spatiotemporal feature fusion matrix is ​​constructed. The dimensions of the spatiotemporal feature fusion matrix include time step, number of feature nodes, and feature dimension. An adaptive topology generation module is deployed in a pre-defined spatiotemporal graph convolutional network. The adaptive topology generation module is used to dynamically adjust the edge weights between feature nodes according to the vehicle type, the driving segment feature parameters, and the braking segment feature parameters. Based on the dynamic edge weight matrix output by the adaptive topology generation module, the spatiotemporal feature fusion matrix is ​​input into the three-layer spatiotemporal convolution kernel of the spatiotemporal graph convolutional network, and spatial graph convolution, temporal convolution and attention weighting are performed in sequence to obtain the convolutional feature matrix. The convolutional feature matrix is ​​then subjected to global pooling and fully connected layer mapping to obtain a comprehensive security index.

[0010] As a further optimization of this disclosure, the adaptive topology generation module adjusts the edge weights between feature nodes using the following formula: ; in, Let be the dynamic edge weight between node i and node j at time t. To pre-define a learnable weight matrix, and Let i and j be the feature vectors of nodes i and j at time t, respectively. Let be the rate of change of vehicle body pitch angle at time t. For vehicle type coding, and These represent the rate of change of acceleration at time t and the maximum rate of change of acceleration, respectively. , These represent the braking response delay at time t and the maximum braking response delay, respectively. Here, e is the spatial attenuation coefficient, and e is a constant. It represents a symmetric difference set.

[0011] As a further optimization of this disclosure, a safety assessment is performed based on the comprehensive safety index and a preset global threshold to obtain a assessment result. This assessment result is used to adjust the bench test parameters, including: The risk level is determined based on the comprehensive safety index and the preset global threshold, and the result is obtained. The risk levels include safe, reduced power operation and forced cooling. When the determination result is that the power is reduced during operation, the drive power is adjusted according to the difference between the comprehensive safety index and the preset global threshold, combined with the speed fluctuation index, and the drive power is adjusted according to the preset power correction rule to obtain the updated drive power parameters. When the determination result is forced cooling, the intensity of the cooling system is dynamically adjusted by the PID controller in combination with the high temperature pixel ratio, and a cooling system control command is generated. The updated drive power parameters and cooling system control commands are sent to the bench control system via the MQTT protocol to generate control commands.

[0012] An electric vehicle operation safety bench testing and evaluation system includes: The data synchronization preprocessing module aligns the acquired raw visual data and raw bench data to obtain synchronized visual data and bench data. The visual data includes wheel rotation trajectory, vehicle pitch angle and brake disc temperature distribution, and the bench data includes vehicle speed, acceleration and brake pedal opening. The multimodal feature hierarchical extraction module extracts features from the visual data to obtain motion state features and component state features, and extracts features from the bench data according to a preset autoencoder network to obtain driving segment feature parameters and braking segment feature parameters; the motion state features include wheel speed change rate and vehicle pitch angle change parameters, and the component state features include the proportion of high-temperature pixels; The spatiotemporal graph convolutional safety assessment module performs feature fusion processing on the motion state features, the component state features, the driving segment feature parameters, and the braking segment feature parameters to obtain a spatiotemporal feature fusion matrix. The spatiotemporal feature fusion matrix is ​​then input into a preset spatiotemporal graph convolutional network for safety assessment processing, and a comprehensive safety index is output. The safety decision and parameter tuning module performs safety judgment processing based on the comprehensive safety index and the preset global threshold to obtain the judgment result, which is used to adjust the bench test parameters.

[0013] An electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor is used to execute the program stored in the memory to implement the electric vehicle operation safety bench test and evaluation method.

[0014] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the electric vehicle operation safety bench testing and evaluation method.

[0015] The beneficial effects of this disclosure are as follows: By aligning the raw visual and bench data, biases from multiple data sources can be eliminated, ensuring that the timestamps of visual data such as wheel rotation trajectories and brake disc temperature distribution are synchronized with bench data such as vehicle speed and acceleration. This provides a spatiotemporal consistency foundation for subsequent feature fusion. Secondly, feature extraction from both visual and bench data allows for the precise capture of key features in electric vehicle operation from both visual and bench perspectives. Motion state features, such as wheel speed change rate and vehicle pitch angle change parameters, reflect the vehicle's dynamic driving state. Driving segment feature parameters and braking segment feature parameters further extract deeper information related to vehicle driving and braking performance from the bench data, enriching the dimensions and levels of features and providing comprehensive and detailed feature support for subsequent safety assessments. Feature fusion of motion state features, component state features, and driving segment feature parameters fully leverages the advantages of different features. Inputting this into a spatiotemporal graph convolutional network for safety assessment fully considers the temporal and spatial correlation of features, enabling a comprehensive, dynamic, and accurate assessment of electric vehicle operational safety. Finally, safety is determined based on the comprehensive safety index and preset global thresholds. The test conditions can be adjusted in a timely manner according to the evaluation results, further optimizing the testing process, improving testing efficiency and accuracy, and ensuring the scientific and effective nature of electric vehicle operation safety testing.

[0016] Compared with traditional electric vehicle operation safety testing methods, this method, by integrating visual perception technology with traditional bench data, not only improves the accuracy of electric vehicle operation safety testing and can more comprehensively and accurately assess the safety status of vehicles under different operating conditions, but also enhances the stability and adaptability of the detection through spatiotemporal graph convolutional networks. It can cope with complex and ever-changing test environments and vehicle operating states, providing an efficient and reliable solution for electric vehicle safety testing. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method in an embodiment of this disclosure; Figure 2 This is a system structure block diagram of an embodiment of this disclosure; Figure 3 This is a block diagram of the device structure in an embodiment of this disclosure. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] like Figure 1 As shown, a bench test and evaluation method for the operational safety of electric vehicles includes the following steps: S1. Align the acquired raw visual data and raw bench data to obtain synchronized visual data and bench data; the visual data includes wheel rotation trajectory, vehicle pitch angle and brake disc temperature distribution, and the bench data includes vehicle speed, acceleration and brake pedal opening. Specifically, raw visual data and raw bench data are collected from the electric vehicle operation safety bench testing environment. Visual data can be collected using devices such as cameras and thermal imagers deployed around the bench. Wheel rotation trajectories can be obtained by capturing the movement of marked points on the wheel surface using a high-speed camera. The vehicle pitch angle can be calculated using tilt sensors installed at specific locations on the vehicle body combined with camera visual information. Brake disc temperature distribution can be generated by real-time temperature monitoring of the brake disc surface using a thermal imager. Bench data can be collected using the vehicle's built-in sensors and bench testing system. However, there are often time discrepancies between the acquisition of raw visual data and raw bench data. For example, the sampling frequencies of the visual sensors and bench sensors may differ. Therefore, time alignment processing, using timestamp calibration and other techniques, can be used to synchronize the raw visual data and raw bench data in time and space, resulting in synchronized visual and bench data.

[0021] S2. Perform feature extraction on the visual data to obtain motion state features and component state features, and perform feature extraction on the bench data according to the preset autoencoder network to obtain drive segment feature parameters and braking segment feature parameters; the motion state features include wheel speed change rate and vehicle body pitch angle change parameters, and the component state features include high temperature pixel ratio; Specifically, the wheel speed change rate reflects the dynamic motion of the wheels, including acceleration, deceleration, and the presence of abnormal states such as slippage. The vehicle pitch angle change parameter reflects the vehicle's attitude changes during driving, which is crucial for assessing the vehicle's handling stability. The high-temperature pixel ratio, as a component status feature, can intuitively reflect the thermal load of key components such as the brake disc, allowing for timely detection of potential overheating risks. Furthermore, an autoencoder network is used to extract features from the bench data, obtaining driving segment feature parameters and braking segment feature parameters. This autoencoder network can automatically learn the inherent features in the data, reducing data dimensionality while retaining key information, and extracting parameters that characterize the vehicle's driving and braking processes. These parameters effectively characterize key information about the vehicle during driving and braking.

[0022] S3. Perform feature fusion processing on the motion state features, component state features, driving segment feature parameters and braking segment feature parameters to obtain a spatiotemporal feature fusion matrix, and input the spatiotemporal feature fusion matrix into a preset spatiotemporal graph convolutional network for safety assessment processing, and output a comprehensive safety index. Specifically, features can be combined into a multi-dimensional matrix, namely a spatiotemporal feature fusion matrix, according to a certain order and rules. This fusion process can fully consider the temporal and spatial correlations of different features, ultimately forming a feature matrix that comprehensively reflects the vehicle's operating status. The spatiotemporal graph convolutional network (SPCRN) is a deep learning model capable of effectively handling the complex relationships between spatiotemporal data. Inputting this spatiotemporal feature fusion matrix into the SPCRN allows for simultaneous convolution operations on both the temporal and spatial dimensions of the matrix. Through multiple layers of convolution operations, it automatically learns the complex relationships and patterns between features, ultimately outputting a comprehensive safety index. This comprehensive safety index quantitatively and intuitively reflects the safety status of the electric vehicle under its current operating condition, providing a scientific and accurate quantitative basis for subsequent safety assessments.

[0023] S4. Based on the comprehensive safety index and the preset global threshold, a safety judgment process is performed to obtain a judgment result, which is used to adjust the bench test parameters.

[0024] Specifically, the preset global threshold is a benchmark value set based on the safety standards and operational requirements of electric vehicles, after extensive experimentation and data analysis. When the overall safety index is lower than the preset global threshold, the vehicle can be judged to have a safety hazard; conversely, if it is higher, the vehicle can be judged to be operating safely. The final judgment result is then obtained. This result not only provides testing personnel with an intuitive safety conclusion but can also be directly used to adjust bench test parameters. For example, based on the judgment result, the bench test conditions can be optimized, the test process adjusted, or the vehicle's control system parameters corrected, thereby further improving the efficiency and accuracy of electric vehicle operational safety testing and ensuring that the safety performance of electric vehicles is fully verified under various operating conditions.

[0025] The above method first aligns the raw visual data and bench data to avoid analytical biases caused by data asynchrony, laying a data foundation for subsequent accurate feature extraction and safety assessment. Secondly, motion state features and component state features are extracted from the visual data, and an autoencoder network is used to extract drive and braking segment feature parameters from the bench data. This allows for the capture of electric vehicle operating information from multiple dimensions, including wheel motion, vehicle posture, component temperature, and vehicle drive and braking states, further enhancing the richness and accuracy of feature representation. Furthermore, the multi-dimensional features are fused to generate a spatiotemporal feature fusion matrix, which is then used for safety assessment via a spatiotemporal graph convolutional network. This fully considers the temporal sequence relationships and spatial correlations of the data, effectively uncovering potential connections between data points and making the safety assessment results more scientific and reasonable. Finally, safety is determined based on the comprehensive safety index and preset global threshold, and bench test parameters are adjusted according to the determination results. This achieves closed-loop control from data processing to evaluation decision-making and parameter optimization. It can not only accurately determine the safe operating status of electric vehicles, but also make bench testing more in line with actual operating needs by dynamically adjusting test parameters, thereby improving the practicality and effectiveness of testing and evaluation, and providing strong support for the accurate assessment and optimization of electric vehicle operating safety.

[0026] In one embodiment, the acquired raw visual data and raw bench data are aligned to obtain synchronized visual data and bench data, including: A synchronous trigger pulse signal is generated through the FPGA hardware triggering mechanism. The synchronous trigger pulse signal is used to control the vision acquisition system and the CAN bus acquisition device to start synchronously and acquire raw vision data and raw bench data respectively. Based on the internal clock of the vision acquisition system and the clock of the CAN bus acquisition device, the acquired raw vision data and raw bench data are timestamped to obtain an aligned timestamp sequence. Based on the aligned timestamp sequence, the original visual data and original bench data are resampled using cubic spline interpolation to obtain the visual data and bench data.

[0027] Specifically, a vision acquisition system can acquire raw visual data such as wheel rotation trajectories, vehicle pitch angles, and brake disc temperature distribution during vehicle operation testing. This raw visual data can intuitively reflect the motion state and thermodynamic condition of key vehicle components. Furthermore, raw bench data such as vehicle speed, acceleration, and brake pedal opening can be collected using CAN bus acquisition devices. This raw bench data reflects the vehicle's performance in terms of dynamics and handling. Due to its high flexibility and speed, the FPGA can precisely control the startup time of multiple devices. Therefore, an FPGA hardware triggering mechanism can be used to generate a synchronous trigger pulse signal. This synchronous trigger pulse signal is used to simultaneously start the vision acquisition system and the CAN bus acquisition device, ensuring that both have the same starting time when acquiring data. In addition, the acquired raw visual data and raw bench data can be timestamped according to the internal clock of the vision acquisition system and the clock of the CAN bus acquisition device to eliminate clock differences between different systems.

[0028] Furthermore, cubic spline interpolation can be used to resample the original visual and rig data. This involves segmenting the data according to the drive-glide-brake cycle, calculating the data value at each timestamp using a preset spline function and target sampling frequency, thus obtaining aligned visual and rig data. This interpolation method ensures that the smoothness and continuity of the data are well preserved during resampling, thereby improving data quality.

[0029] In one embodiment, feature extraction is performed on visual data to obtain motion state features and component state features, including: Based on the image data corresponding to the wheel rotation trajectory, the pixel displacement difference between adjacent frames is calculated using the optical flow method, and the optical flow equation is solved using the Lucas-Kanade algorithm to obtain the wheel rotation speed change rate. Based on the binocular vision data corresponding to the vehicle pitch angle, the coordinates of the preset marker points on the vehicle are matched by a 3D reconstruction algorithm, and the vehicle pitch angle change parameters are output. Based on the brake disc temperature distribution, the U-Net segmentation network is used to segment the infrared image corresponding to the brake disc temperature distribution, calculate the pixel ratio of abnormal high temperature areas, and obtain the high temperature pixel ratio.

[0030] Specifically, optical flow is a computer vision technique used to estimate pixel motion in image sequences. By calculating the displacement difference between pixels in adjacent frames, optical flow can capture the motion trajectory of pixels as a wheel rotates. The Lucas-Kanade algorithm is an implementation of optical flow, which estimates the motion vector of pixels by solving the optical flow equation, thereby achieving real-time tracking of wheel rotation speed. Specifically, this algorithm first selects feature points in the initial frame, such as key points on the wheel contour. Then, in subsequent frames, it finds the corresponding positions of these feature points, uses gradient information and temporal differences to establish and solve the optical flow equation, obtaining the motion vector of each feature point. By analyzing these motion vectors, the rate of change of wheel rotation speed can be calculated, reflecting the wheel's rotational state, including acceleration, deceleration, and the presence of abnormal vibrations, providing important information for evaluating the vehicle's driving and braking performance.

[0031] Indicatively, this binocular vision data can be acquired by simultaneously capturing images of the vehicle body from different angles using two cameras. Pre-set markers on the vehicle body are specific locations pre-positioned on the vehicle and serve as reference points for 3D reconstruction. The 3D reconstruction algorithm analyzes the two-dimensional image coordinates of these markers in the binocular vision data, combining this with the camera's intrinsic and extrinsic parameters, such as focal length, optical center coordinates, and the relative position between the cameras, to calculate the markers' coordinates in 3D space using triangulation principles. By comparing and analyzing the 3D coordinates at different times, the vehicle's pitch angle variation parameter can be calculated. This pitch angle variation parameter reflects the degree of vertical tilt of the front or rear of the vehicle body caused by acceleration or braking during driving.

[0032] Furthermore, visual data on the temperature distribution of the brake disc can be acquired using an infrared thermal imaging camera. This visual data can reflect the temperature changes of the brake disc in real time during operation. The U-Net segmentation network is an image segmentation model based on a convolutional neural network. Illustratively, the acquired infrared image can first be normalized and enhanced to improve image quality and segmentation results. Then, the preprocessed image is input into the trained U-Net network. This network can extract multi-scale features of the image through multiple convolutions, downsampling, upsampling, and skip connections, outputting the segmentation result, dividing the image into normal temperature regions and abnormally high temperature regions. Finally, the proportion of high-temperature pixels can be obtained by calculating the ratio of pixels in the abnormally high-temperature region to the total number of pixels in the brake disc image. This high-temperature pixel ratio can intuitively reflect the thermal load of the brake disc. When it exceeds a certain threshold, it indicates that the brake disc is at risk of overheating, thus affecting braking performance and safety, providing crucial information for timely detection and handling of braking system faults.

[0033] In one embodiment, feature extraction is performed on the bench data using an autoencoder network to obtain drive segment feature parameters and braking segment feature parameters, including: An autoencoder network based on gated recurrent units and temporal attention mechanism is constructed. The encoder of the autoencoder network includes an LSTM unit, a GRU gate module, and an attention weight calculation module. Based on vehicle speed and acceleration, the GRU module and LSTM unit are used to perform time series correlation filtering and time series encoding respectively to obtain the LSTM hidden state sequence. The attention weight calculation module is used to weight the LSTM hidden state sequence to obtain the weighted driving time series features. Time series analysis was performed on the weighted driving time series characteristics to calculate the acceleration change rate and velocity fluctuation index, thereby obtaining the characteristic parameters of the driving segment. The opening gradient sequence is calculated based on the brake pedal opening degree. The sequence is then processed by the GRU module and the LSTM unit for temporal correlation filtering and gradient encoding, respectively, to obtain the LSTM gradient hidden state sequence. The LSTM gradient hidden state sequence is weighted by the attention weight calculation module to obtain the weighted braking gradient features; Gradient analysis is performed on the weighted braking gradient characteristics to calculate the braking response delay and opening change gradient, thereby obtaining the characteristic parameters of the braking segment.

[0034] Specifically, vehicle speed and acceleration data are sequentially input into the GRU module and the LSTM unit. The GRU module, through its update and reset gates, filters the time-series data, retaining important information relevant to the current moment and removing irrelevant information, thus achieving time-series correlation filtering. This filters out strongly correlated segments such as rapid acceleration and suppresses weakly correlated noise such as idling and coasting. The processed data is then input into the LSTM unit, which encodes the time series, captures long-term dependencies, and obtains the LSTM hidden state sequence. This sequence contains key feature information of vehicle speed and acceleration over time. This sequence is then input into the attention weight calculation module, which dynamically assigns different weights based on the importance of features at each moment in the sequence, highlighting important features and weakening less important ones. After weighted processing, a weighted driving time-series feature is obtained, which focuses more on the key parts of the vehicle speed and acceleration data closely related to the vehicle's driving state. Based on this feature, the mean absolute difference of weighted acceleration between adjacent time points can be calculated using a sliding window to obtain the rate of acceleration change, or the speed fluctuation index can be obtained by calculating the standard deviation of the weighted vehicle speed sequence.

[0035] Specifically, calculating the gradient of the original brake pedal opening sequence can more accurately reflect the aggressiveness of the braking operation. Similarly, by sequentially passing the opening gradient sequence through the GRU module, LSTM unit, and attention weight calculation module for temporal correlation filtering, gradient encoding, and weighting, effective segments of brake pedal opening gradient change can be selected, cross-time correlations in bench data can be captured, and key time step features such as rapid acceleration and rapid brake pedal depressing can be enhanced. Finally, based on the weighted brake gradient features, the braking response delay can be calculated by the time difference between the gradient abrupt change moment and the peak braking pressure moment, and the opening change gradient can be obtained by calculating the mean of the weighted opening gradient sequence. This gradient reflects the rate of change of brake pedal opening and can be used to evaluate the smoothness of braking operation and the sensitivity of the braking system.

[0036] In one embodiment, feature fusion processing is performed based on motion state features, component state features, driving segment feature parameters, and braking segment feature parameters to obtain a spatiotemporal feature fusion matrix. This spatiotemporal feature fusion matrix is ​​then input into a spatiotemporal graph convolutional network for safety assessment, outputting a comprehensive safety index, including: Based on motion state characteristics, component state characteristics, driving segment characteristic parameters, and braking segment characteristic parameters, a spatiotemporal feature fusion matrix is ​​constructed. The dimensions of the spatiotemporal feature fusion matrix include time step, number of feature nodes, and feature dimension. An adaptive topology generation module is deployed in the spatiotemporal graph convolutional network. The adaptive topology generation module is used to dynamically adjust the edge weights between feature nodes according to the vehicle type, driving segment feature parameters and braking segment feature parameters. Based on the dynamic edge weight matrix output by the adaptive topology generation module, the spatiotemporal feature fusion matrix is ​​input into the three-layer spatiotemporal convolution kernel of the spatiotemporal graph convolutional network, and spatial graph convolution, temporal convolution and attention weighting are performed in sequence to obtain the convolutional feature matrix. The convolutional feature matrix is ​​then subjected to global pooling and fully connected layer mapping to obtain the comprehensive security index.

[0037] Specifically, the time step represents the number of intervals at which features are sampled in the time series, including multiple consecutive moments during vehicle operation. The number of feature nodes corresponds to different parts of the vehicle or different types of feature sources, where each feature node represents a specific feature measurement point or feature type. Feature dimension refers to the number of specific feature parameters contained in each feature node, reflecting the multi-dimensional nature of the feature. By constructing a fusion matrix including the time step, the number of feature nodes, and the feature dimension, features from different data sources and of different types can be uniformly integrated, providing structured input data for subsequent spatiotemporal graph convolutional network processing.

[0038] Secondly, different types of vehicles have different structures and operating characteristics. For example, sedans, SUVs, and trucks differ in body mass, center of gravity distribution, and powertrain. The adaptive topology generation module can adjust the correlation between feature nodes according to vehicle type. For instance, for SUVs with a high center of gravity, the correlation weight between the body pitch angle change parameter and other feature nodes can be increased accordingly. Furthermore, driving segment feature parameters such as acceleration change rate and speed fluctuation index, and braking segment feature parameters such as braking response delay and braking angle change gradient, can reflect the dynamic characteristics of the vehicle under different operating conditions. Therefore, the adaptive topology generation module can dynamically adjust the weights between feature nodes based on changes in these parameters. For example, when the acceleration change rate is large, the correlation weight between the wheel speed change rate and other feature nodes (such as the body pitch angle change parameter) can be enhanced because the vehicle's driving state is more intense. Further, the spatiotemporal feature fusion matrix can be used as the node feature matrix of the graph. Based on the adjacency relationship constrained by the dynamic edge weight matrix output by this module, convolution is performed through a spatiotemporal graph convolutional network.

[0039] The adaptive topology generation module adjusts the edge weights between feature nodes using the following formula: ; in, Let be the dynamic edge weight between node i and node j at time t. To pre-define a learnable weight matrix, and Let i and j be the feature vectors of nodes i and j at time t, respectively. Let be the rate of change of vehicle body pitch angle at time t. For vehicle type coding, and These represent the rate of change of acceleration at time t and the maximum rate of change of acceleration, respectively. , These represent the braking response delay at time t and the maximum braking response delay, respectively. Here, e is the spatial attenuation coefficient, and e is a constant. It represents a symmetric difference set.

[0040] Specifically, the spatiotemporal feature fusion matrix can be input into the three-layer spatiotemporal convolution kernel of the spatiotemporal graph convolutional network. Based on the inter-node relationships defined by the dynamic edge weight matrix, the features of each node are aggregated and updated, i.e., spatial convolution is performed. This convolution process, by considering the feature information of adjacent nodes, further enhances the spatial correlation and globality of the features, and can capture the mutual influence between features of different parts of the vehicle. Furthermore, by performing convolution operations on the spatiotemporal feature fusion matrix in the time series dimension, the evolution pattern of features during vehicle operation can be captured. Subsequently, an attention mechanism can be introduced to weight the convolved features, enabling the network to automatically learn the importance of different time steps and different feature nodes. Finally, a global pooling operation is performed on the feature matrix after spatiotemporal graph convolution processing to compress the spatial and temporal dimensions of the feature matrix, reduce the dimensionality of the data, and extract key feature information. The pooled features are then mapped to a scalar value, i.e., the comprehensive safety index, through a fully connected layer. This comprehensive safety index is a quantitative assessment result of the electric vehicle's operational safety status, reflecting the overall safety level of the vehicle under its current operating condition.

[0041] In one embodiment, a safety assessment is performed based on a comprehensive safety index and a preset global threshold to obtain a assessment result. This assessment result is used to adjust the bench test parameters, including: The risk level is determined based on the comprehensive safety index and the preset global threshold. The risk levels include safe, reduced power operation and forced cooling. When the determination result is that the power is reduced during operation, the drive power is adjusted according to the difference between the comprehensive safety index and the preset global threshold, combined with the speed fluctuation index, and the drive power is adjusted according to the preset power correction rule to obtain the updated drive power parameters. When the determination result is forced cooling, the intensity of the cooling system is dynamically adjusted by the PID controller based on the proportion of high-temperature pixels, and a cooling system control command is generated. The updated drive power parameters and cooling system control commands are sent to the bench control system via the MQTT protocol to generate control commands.

[0042] Specifically, when the overall safety index is higher than a preset global threshold, the vehicle can be determined to be in a safe state. When the overall safety index is lower than the preset global threshold but higher than another lower warning threshold, it can be determined to be in a power-reduced operation state. In this case, although the vehicle still faces certain risks, it can maintain safe operation by reducing power. Furthermore, when the overall safety index is lower than the warning threshold, it can be determined to be in a forced cooling state, indicating that certain components of the vehicle, such as the brake discs, may be overheating and require immediate cooling measures to avoid potential safety risks.

[0043] Indicatively, when the determination result is reduced power operation, the drive power can be adjusted according to a preset power correction rule based on the difference between the comprehensive safety index and the preset global threshold, combined with the speed fluctuation index. The difference between the comprehensive safety index and the preset global threshold reflects the degree to which the vehicle's current operating state deviates from the safe range. The larger the difference, the higher the risk. The speed fluctuation index is one of the characteristic parameters of the drive segment, reflecting the stability of the vehicle's speed during driving. When speed fluctuations are large, it indicates that there may be unstable factors in the vehicle's drive system. Based on the above difference and speed fluctuation index, the updated drive power parameters can be calculated according to a preset rule. This rule can be a complex functional relationship or an empirical formula based on experimental data, ensuring that while reducing power, safety risks are avoided and the impact on normal vehicle operation is minimized.

[0044] Specifically, the proportion of high-temperature pixels is one of the component status characteristics, reflecting the temperature of key components such as brake discs. By monitoring the proportion of high-temperature pixels, the thermal load of components can be understood in real time, providing a direct basis for adjusting the intensity of the cooling system. The PID controller is a closed-loop control algorithm that can perform proportional, integral, and derivative control based on the deviation. In this embodiment, when the determination result is forced cooling, based on the PID controller, the working intensity of the cooling system can be dynamically adjusted according to the proportion of high-temperature pixels, such as adjusting the speed of the cooling fan or the flow rate of the coolant, to quickly reduce the component temperature. Finally, the updated drive power parameters and cooling system control commands can be sent to the bench control system via the MQTT protocol to generate control commands. After receiving the control commands, the bench control system can adjust the vehicle's operating state according to the command content, such as reducing the drive power or increasing the intensity of the cooling system, thereby achieving effective intervention in the safe operation of electric vehicles and ensuring that the vehicle is always in a safe and controllable state during the test.

[0045] like Figure 2 As shown, embodiments of this disclosure provide an electric vehicle operation safety bench testing and evaluation system, including: The data synchronization preprocessing module aligns the acquired raw visual data and raw bench data to obtain synchronized visual data and bench data. The visual data includes wheel rotation trajectory, vehicle pitch angle and brake disc temperature distribution, and the bench data includes vehicle speed, acceleration and brake pedal opening. The multimodal feature hierarchical extraction module extracts features from the visual data to obtain motion state features and component state features, and extracts features from the bench data according to a preset autoencoder network to obtain driving segment feature parameters and braking segment feature parameters; the motion state features include wheel speed change rate and vehicle pitch angle change parameters, and the component state features include the proportion of high-temperature pixels; The spatiotemporal graph convolutional safety assessment module performs feature fusion processing on the motion state features, the component state features, the driving segment feature parameters, and the braking segment feature parameters to obtain a spatiotemporal feature fusion matrix. The spatiotemporal feature fusion matrix is ​​then input into a preset spatiotemporal graph convolutional network for safety assessment processing, and a comprehensive safety index is output. The safety decision and parameter tuning module performs safety judgment processing based on the comprehensive safety index and the preset global threshold to obtain the judgment result, which is used to adjust the bench test parameters.

[0046] The implementation process of the functions and roles of each module in the above system is detailed in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0047] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0048] See Figure 3 The electronic device provided in the embodiments of this disclosure includes a processor 1110, a communication interface 1120, a memory 1130 and a communication bus 1140, wherein the processor 1110, the communication interface 1120 and the memory 1130 communicate with each other through the communication bus 1140. Memory 1130 is used to store computer programs; The processor 1110, when executing the program stored in the memory 1130, implements the above-described electric vehicle operation safety bench test and evaluation method. The aforementioned communication bus 1140 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus 1140 can be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, it is represented by only one thick line in the figure, but this does not indicate that there is only one bus or one type of bus.

[0049] The communication interface 1120 is used for communication between the above-mentioned electronic device and other devices.

[0050] The memory 1130 may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1130 may also be at least one storage device located remotely from the aforementioned processor 1110.

[0051] Embodiments of this disclosure also provide a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when executed by a processor, implements the electric vehicle operation safety bench testing and evaluation method as described above.

[0052] The embodiments described above are merely examples of several implementations of this disclosure, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent disclosure. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this disclosure, and these modifications and improvements all fall within the protection scope of this disclosure.

Claims

1. A bench test and evaluation method for the operational safety of electric vehicles, characterized in that, Includes the following steps: The acquired raw visual data and raw bench data are aligned to obtain synchronized visual data and bench data; the visual data includes wheel rotation trajectory, vehicle pitch angle and brake disc temperature distribution, and the bench data includes vehicle speed, acceleration and brake pedal opening. Feature extraction is performed on the visual data to obtain motion state features and component state features. Feature extraction is also performed on the bench data according to a preset autoencoder network to obtain drive segment feature parameters and braking segment feature parameters. The motion state features include wheel speed change rate and vehicle pitch angle change parameters. The component state features include the proportion of high-temperature pixels. The motion state features, component state features, driving segment feature parameters, and braking segment feature parameters are fused to obtain a spatiotemporal feature fusion matrix. The spatiotemporal feature fusion matrix is ​​then input into a preset spatiotemporal graph convolutional network for safety assessment and outputs a comprehensive safety index. Based on the comprehensive safety index and the preset global threshold, a safety assessment is performed to obtain the assessment result, which is used to adjust the bench test parameters.

2. The method for bench testing and evaluation of electric vehicle operation safety according to claim 1, characterized in that, The acquired raw visual data and raw test bench data are aligned to obtain synchronized visual data and test bench data, including: A synchronous trigger pulse signal is generated through the FPGA hardware triggering mechanism. The synchronous trigger pulse signal is used to control the vision acquisition system and the CAN bus acquisition device to start synchronously and acquire raw vision data and raw bench data respectively. Based on the internal clock of the vision acquisition system and the clock of the CAN bus acquisition device, the acquired raw vision data and raw bench data are timestamped to obtain an aligned timestamp sequence. Based on the aligned timestamp sequence, the original visual data and original test bench data are resampled using cubic spline interpolation to obtain synchronized visual data and test bench data.

3. The method for bench testing and evaluation of electric vehicle operation safety according to claim 1, characterized in that, Feature extraction is performed on the visual data to obtain motion state features and component state features, including: Based on the image data corresponding to the wheel rotation trajectory, the pixel displacement difference between adjacent frames is calculated using the optical flow method, and the optical flow equation is solved using the Lucas-Kanade algorithm to obtain the wheel rotation speed change rate. Based on the binocular vision data corresponding to the vehicle pitch angle, the coordinates of the preset marker points on the vehicle are matched by a three-dimensional reconstruction algorithm, and the vehicle pitch angle change parameters are output. Based on the brake disc temperature distribution, the U-Net segmentation network is used to segment the infrared image corresponding to the brake disc temperature distribution, calculate the pixel ratio of abnormal high temperature areas, and obtain the high temperature pixel ratio.

4. The method for bench testing and evaluation of electric vehicle operation safety according to claim 1, characterized in that, Feature extraction is performed on the test bench data based on a preset autoencoder network to obtain drive segment feature parameters and braking segment feature parameters, including: An autoencoder network based on gated recurrent units and temporal attention mechanism is constructed. The encoder of the autoencoder network includes an LSTM unit, a GRU gate module, and an attention weight calculation module. Based on vehicle speed and acceleration, the GRU module and the LSTM unit are sequentially used to perform temporal correlation filtering and time series encoding to obtain the LSTM hidden state sequence. The attention weight calculation module is then used to weight the LSTM hidden state sequence to obtain the weighted driving temporal features. Time series analysis is performed on the weighted driving time series features to calculate the acceleration change rate and speed fluctuation index, and driving segment feature parameters are obtained. The opening gradient sequence is calculated based on the brake pedal opening, and the time series correlation filtering and gradient encoding are performed sequentially through the GRU module and the LSTM unit to obtain the LSTM gradient hidden state sequence. The attention weight calculation module is used to weight the LSTM gradient hidden state sequence to obtain the weighted braking gradient features. Gradient analysis is performed on the weighted braking gradient characteristics to calculate the braking response delay and opening change gradient, thereby obtaining the braking segment characteristic parameters.

5. The method for bench testing and evaluation of electric vehicle operation safety according to claim 1, characterized in that, The motion state features, component state features, driving segment feature parameters, and braking segment feature parameters are fused to obtain a spatiotemporal feature fusion matrix. This spatiotemporal feature fusion matrix is ​​then input into a preset spatiotemporal graph convolutional network for security assessment, outputting a comprehensive security index, including: Based on the motion state features, component state features, driving segment feature parameters, and braking segment feature parameters, a spatiotemporal feature fusion matrix is ​​constructed. The dimensions of the spatiotemporal feature fusion matrix include time step, number of feature nodes, and feature dimension. An adaptive topology generation module is deployed in a pre-defined spatiotemporal graph convolutional network. The adaptive topology generation module is used to dynamically adjust the edge weights between feature nodes according to the vehicle type, the driving segment feature parameters, and the braking segment feature parameters. Based on the dynamic edge weight matrix output by the adaptive topology generation module, the spatiotemporal feature fusion matrix is ​​input into the three-layer spatiotemporal convolution kernel of the spatiotemporal graph convolutional network, and spatial graph convolution, temporal convolution and attention weighting are performed in sequence to obtain the convolutional feature matrix. The convolutional feature matrix is ​​then subjected to global pooling and fully connected layer mapping to obtain a comprehensive security index.

6. The method for bench testing and evaluation of electric vehicle operation safety according to claim 5, characterized in that, The adaptive topology generation module adjusts the edge weights between feature nodes using the following formula: ； in, Let be the dynamic edge weight between node i and node j at time t. To pre-define a learnable weight matrix, and Let i and j be the feature vectors of nodes i and j at time t, respectively. Let be the rate of change of vehicle body pitch angle at time t. For vehicle type coding, and These represent the rate of change of acceleration at time t and the maximum rate of change of acceleration, respectively. , These represent the braking response delay at time t and the maximum braking response delay, respectively. Here, e is the spatial attenuation coefficient, and e is a constant. It represents a symmetric difference set.

7. The method for bench testing and evaluation of electric vehicle operation safety according to claim 1, characterized in that, Based on the comprehensive safety index and the preset global threshold, a safety assessment is performed to obtain a assessment result. The assessment result is used to adjust the bench test parameters, including: The risk level is determined based on the comprehensive safety index and the preset global threshold, and the result is obtained. The risk levels include safe, reduced power operation and forced cooling. When the determination result is that the power is reduced during operation, the drive power is adjusted according to the difference between the comprehensive safety index and the preset global threshold, combined with the speed fluctuation index, and the drive power is adjusted according to the preset power correction rule to obtain the updated drive power parameters. When the determination result is forced cooling, the intensity of the cooling system is dynamically adjusted by the PID controller in combination with the high temperature pixel ratio, and a cooling system control command is generated. The updated drive power parameters and cooling system control commands are sent to the bench control system via the MQTT protocol to generate control commands.

8. A bench testing and evaluation system for the operational safety of electric vehicles, characterized in that, include: The data synchronization preprocessing module aligns the acquired raw visual data and raw bench data to obtain synchronized visual data and bench data. The visual data includes wheel rotation trajectory, vehicle pitch angle and brake disc temperature distribution, and the bench data includes vehicle speed, acceleration and brake pedal opening. The multimodal feature hierarchical extraction module extracts features from the visual data to obtain motion state features and component state features, and extracts features from the bench data according to a preset autoencoder network to obtain driving segment feature parameters and braking segment feature parameters; the motion state features include wheel speed change rate and vehicle pitch angle change parameters, and the component state features include the proportion of high-temperature pixels; The spatiotemporal graph convolutional safety assessment module performs feature fusion processing on the motion state features, the component state features, the driving segment feature parameters, and the braking segment feature parameters to obtain a spatiotemporal feature fusion matrix. The spatiotemporal feature fusion matrix is ​​then input into a preset spatiotemporal graph convolutional network for safety assessment processing, and a comprehensive safety index is output. The safety decision and parameter tuning module performs safety judgment processing based on the comprehensive safety index and the preset global threshold to obtain the judgment result, which is used to adjust the bench test parameters.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor for executing a program stored in a memory to implement the electric vehicle operation safety bench testing and evaluation method according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the electric vehicle operation safety bench testing and evaluation method according to any one of claims 1-7.

Images